A deep reinforcement learning framework for multi-dimensional adaptability in building operations

Mission-critical buildings, like hospitals, must adapt continuously to increasing service demands despite constrained resources. Adaptation strategies, including spatial reconfiguration, workforce adjustments, and operational changes, are often implemented in isolation, neglecting interdependencies, varying timescales, and conflicting objectives, leading to inefficiencies. This study introduces a Deep Reinforcement Learning (DRL) framework to explore and optimize multi-dimensional adaptation strategies across modalities (spatial reconfiguration, workforce adjustments, and operational changes), objectives (spatial efficiency, user experience, and operational performance), and implementation timescales (from minutes to years). The framework integrates (a) a narrative-based formalism for modeling operations in mission-critical facilities, (b) an action-masked Proximal Policy Optimization (PPO) algorithm that coordinates concurrent resource and activity allocations under hard sequencing constraints, and (c) a post hoc clustering and ranking method with Pareto-front analysis to extract interpretable adaptation strategies and multi-objective trade-offs. In a cardiac catheterization lab case study, we evaluated 48 multi-dimensional adaptation scenarios and clustered outcomes into 17 macro-level strategies, identifying three that best balance space utilization, patient experience, and operational efficiency. The PPO agent reached 80% of optimal performance after 3.5M training steps and near-optimal performance after 9.8M steps, outperforming tabular Q-learning and an action-masked untrained policy baseline. This work advances decision-making in building operations, enabling resilient, multi-timescale responses to evolving service demands in mission-critical environments.